The world's idle GPUs. In one pool.

Cheaper than big cloud, straighter than RunPod: rent a GPU with per-second billing, a hard budget cap and the promise that you don't pay for seconds that don't work — or put your own idle card up for rent.

No commitment: you pay by the second and stop whenever you want. The product is in Phase 0 — there is no public sign-up flow yet, and the console runs on a local development address.

Marketplace — example view illustrative offers · EU + TR
RTX 4090 Istanbul, TR 4.9 $0.42/hr
H100 SXM Helsinki, FI 5.0 $2.19/hr
RTX A6000 Ankara, TR 4.6 $0.88/hr
L40S Warsaw, PL 4.7 $0.79/hr
RTX 3090 Bursa, TR 4.5 $0.21/hr
The scorecard score drives placement. The rows above are not a live pool; they are an illustrative example of the interface.

You don't pay for seconds that don't work

If a node fails its health check, billing stops that same second and your job moves to another host from its checkpoint. On any day you were charged for a broken pod, we refund the difference — without being asked.

Hard budget cap

Set your monthly limit; when the ceiling is reached, jobs pause safely. Every line item — including the storage charge on a stopped pod — sits on one screen. "Surprise invoice" is not a category here.

Per-second billing

Not per minute, and certainly not per hour: you pay for the seconds you use. Start-up takes under two minutes, shutdown is one click, commitment is zero.

These three points are the product's commitment, not a report of past performance: the metering and the health gate are one of Phase 0's delivery slices and are not running in production yet. Component-by-component status: Status.

Starting templates

Ready-made pods: pick one, start

Nobody wants an empty GPU; they want something that runs. A template is one image plus sane defaults — disk, ports, a recommended GPU. Pick one here and the console opens its rent dialog with that template already selected. Per-second billing starts when the pod does.

Jupyter + PyTorch

Experiment and prototype: from a notebook cell to a training script, without changing machines.

RTX 4090min 24 GB VRAM

vLLM · OpenAI-compatible server

Publish a model as an API; on the client side only base_url changes.

A100 80GBmin 48 GB VRAM

Ollama · your own chat

Pull an open model and run it with a single command; the pod answers on an endpoint that is yours.

RTX 4090min 24 GB VRAM

Axolotl · fine-tuning

LoRA, QLoRA and full fine-tune recipes; checkpoints land on a volume that outlives the pod.

A100 80GBmin 48 GB VRAM

ComfyUI · image generation

SD/SDXL/Flux workflows on a node-based interface. Preparing: the template package is not fixed yet — today the same work runs through the custom-image flow, you just supply the image.

preparingRTX 4090min 24 GB VRAM

Whisper · speech transcription

Transcription, subtitles and batch audio processing. Preparing: the template package is not fixed yet — today it runs through the custom-image flow, you just supply the image.

preparingRTX 3090min 16 GB VRAM

Blender · render

3D render, animation and architectural visualization. Preparing: the template package is not fixed yet — today it runs through the custom-image flow, you just supply the image.

preparingRTX 4090min 24 GB VRAM

Agent runtime

Tool-calling LLM agents and long-running jobs. Preparing: the template package is not fixed yet — today it runs through the custom-image flow, you just supply the image.

preparingRTX 4090min 24 GB VRAM

Custom image

Your own container: paste the registry address, set the disk and the ports, start. Every workload the list above does not cover lands here.

any GPUVRAM: your call

Four of the templates run today with public, official images and the custom image always works; the other four are marked preparing — their package is not fixed yet, so their Start button hands you the custom-image flow and you supply the image. Nothing here carries a "ready" label before it is ready. Image tags are deliberately not printed on this page: tags move, and a stale page is worse than no page — the exact image sits in the console's rent dialog. Full list: Model Catalog.

Workloads

What can you run?

The list is short on purpose: it covers the work people actually run on a rented GPU. Where a template is still preparing, the work itself is possible today anyway — with your own image. Nothing below promises more than that.

Notebook and prototyping — Jupyter + PyTorchready · dev
Serving a model as an API — vLLM, Ollamaready · inference
Fine-tuning: LoRA, QLoRA, full runs — Axolotlready · training
Everything else, in your own container — Custom imageready · dev
Image generation — ComfyUIpreparing · today via custom image
Speech transcription and subtitles — Whisperpreparing · today via custom image
3D render and animation — Blenderpreparing · today via custom image
Tool-calling agents — Agent runtimepreparing · today via custom image
Multi-node cluster trainingnot in the catalog · Clusters, after Phase 2
Whichever template you pick: per-second billing, a hard budget cap, and you don't pay for seconds that don't work.
The scorecard picks the machine: the template says what runs, the scorecard says where it runs — price × reliability × network.
Region pinning still holds: a template does not move your data — the pod stays in the country you pinned it to.

By workload, in more depth: fine-tuning · inference · image & video. Console links point to a local development address.

For people who live in code

The console is good. The API is better.

One API key to search offers, spin up a pod and open an endpoint. A CLI, a Python SDK and OpenAI-compatible endpoints exist from day one — they live inside your infrastructure code, not in browser tabs.

The scorecard is in the API: you pick the best-scoring machine, not the cheapest one — price × reliability × network.
OpenAI-compatible: change base_url, leave the rest of your code alone.
Terraform & SkyPilot support is on the way — come as you are, without breaking your current workflow.
CLIPythoncurl
# Rank the best offers by score — output is illustrative $ kaldera offers GPU QTY REGION $/HR SCORE RANK RTX 4090 1 Istanbul 0.42 4.9 0.90 RTX 4090 4 Frankfurt 1.58 4.8 0.61 H100 SXM 1 Helsinki 2.19 5.0 0.55 # Rent the one you like, by the second $ kaldera pod create --offer ofr-8817 \ --image pytorch/2.4-cuda12.4 ✓ pod ready → ssh root@fra1.kaldera.ai -p 40122

Serverless Inference

Give a model, get a URL.

Paste a Hugging Face model name and get a production endpoint that scales to zero, with a cold-start target under 2 seconds. No requests, no charge — the marketplace carries the cost of the waiting GPU, not you.

< 2 s

Cold-start target: image pre-pull plus a model cache on the host's NVMe. No user stuck in a queue, no endpoint burning money.

scale to 0

When traffic stops the endpoint sleeps, and so does its bill. It wakes on the first request — no standby replica required.

drop-in

OpenAI-compatible API: bring your current client, your LangChain setup, your agent as they are. The only line that changes is base_url.

Serverless inference belongs to Phase 2: there is no endpoint you can open today, and the numbers above are targets — not measured results. Detail: Serverless Inference.

Sovereign Region

Your data stays where you decide it stays.

Pin your pod to a country; your data does not cross that border. We work through EU and Turkish legal entities — we are not a structure from which data can be demanded under the US CLOUD Act. The AI Act is in force and KVKK (Turkish personal data protection law) is on the table: your compliance team's sign-off is the easy part.

Istanbul · Ankara · BursaTurkey — KVKK resident
FrankfurtGermany — EU region
HelsinkiFinland — EU region
AmsterdamNetherlands — EU region
WarsawPoland — EU region
Next in lineBaku · Sofia · Athens
KVKKGDPREU AI ActRegion pinningSOC 2 Type I — in progress

The list above is of targeted regions; being able to select a region does not mean a machine is available there at that moment. SOC 2 Type I has not been obtained — the process stands as a goal to be started in Phase 3. Detail: Sovereignty.

Pricing

Marketplace pricing: you pay for the GPU, not the middleman.

Hosts set the prices and competition pushes them down; we take a 15% commission and we do not hide it. Below are example starting prices from today's pool — the live ones are in the console.

GPUVRAMTypical useStarting price
RTX 309024 GBStable Diffusion, small LLMs$0.21/hr
RTX 409024 GBFine-tuning, image generation$0.42/hr
L40S48 GBInference, video$0.79/hr
RTX A600048 GBLong context, render$0.88/hr
A100 PCIe80 GBTraining, 70B inference$1.32/hr
H100 SXM80 GBSerious training runs$2.19/hr

These are example prices; on the marketplace they move second by second with supply and demand. Storage and traffic are separate, visible line items — tracked with a live counter on the billing screen.

Become a Host

Don't let your card sit idle: 85% is yours.

A rig left over from mining, your studio's render farm or your university cluster — install one binary and join the pool without opening an inbound port. Commission is 15% (the industry is known to sit in the 20–25% band), payouts are weekly, and the scorecard is public: a good host gets more work.

The host agent, machine registration and weekly payouts go live in Phase 1; there is no open application form today.

Example earnings — 4× RTX 4090
Your list price4 × $0.42/hr
Assumed utilization60%
Monthly gross$725.76
Kaldera commission (15%)−$108.86
Monthly net, electricity excluded$616.90

Illustrative calculation: the price and the utilization are assumptions, not measured earnings.

Frequently Asked

Can production work be trusted to a marketplace GPU?

That is exactly what the scorecard is for: every host is benchmarked continuously, its score is public, and critical work is not placed on a low-scoring machine. For enterprise work there is a "Verified" tier with on-site audits and an SLA. And the guarantee holds either way: seconds that don't work are not billed.

How are you different from RunPod and other GPU marketplaces?

Three things: a broken node does not run up a bill (automatic detection + refund + the job moves), the invoice sits on one screen with a hard cap (the storage charge on a stopped pod is no surprise), and our EU + TR legal structure lets us pin data to a region — we do not fall under the US CLOUD Act.

Is my data safe on the host's disk?

Secrets never land on the host's disk; images are signed, sensitive workloads are placed only on the Verified tier, and your persistent data sits in encrypted, S3-compatible storage. The host cannot see inside your container.

Does my job crash when it hits the budget cap?

It does not: the job is paused safely, its checkpoint is kept and a notification reaches you. Raise the limit and continue from the second you left off. The cap is a brake, not a guillotine.

Get started

Let infrastructure be our job; yours is the model.

The console and the API contract were written in Phase 0; there is no public sign-up flow yet. What there is to look at today is the code and the roadmap.

Console links point to the console domain; when you open the site locally they are rewritten to localhost:3000.